Goodhart's Law in the Plant Room

The monthly energy report says the chiller plant averaged 0.62 kW/RT in June. Better than design. Meanwhile the front desk has logged eleven complaints about a warm lobby, the level-three AHU has been running flat out since Raya, and the engineer who compiles the report has a line in his appraisal form that reads "plant efficiency ≤ 0.65 kW/RT." Nobody has lied. Yet the number and the building disagree. There is a half-century-old name for what happened here: Goodhart's Law.

A footnote from 1975

Charles Goodhart, then an adviser at the Bank of England, buried the idea in a 1975 paper called Problems of Monetary Management: The UK Experience, published by the Reserve Bank of Australia. His original wording is drier than the version you know: "any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes". He was talking about monetary aggregates: the moment the Bank of England targeted one, its once-stable relationship with the economy dissolved.

The sharp version, "when a measure becomes a target, it ceases to be a good measure," is not Goodhart's at all. It belongs to the anthropologist Marilyn Strathern, writing in 1997 about audit culture in British universities. Most essays hand her line to him; she deserves the credit. Donald T. Campbell had reached the same conclusion independently in 1979: the more a quantitative indicator is used for decision-making, the more it distorts the process it was meant to monitor.

The pattern is much older than the names. In 1902, French colonial authorities in Hanoi paid a bounty per rat tail to fight plague. Residents cut the tails off live rats and released them to breed more rats; rat-farming and tail-smuggling networks appeared. Historian Michael G. Vann found it all in the colonial archives. (The Soviet nail factory that made a few giant nails when targeted by weight is, by contrast, a parable traced to a satirical cartoon. The rats are real.)

And it scales. Wells Fargo's "eight is great" cross-selling target produced roughly 3.5 million potentially fake accounts and a $100 million CFPB fine in 2016. Volkswagen wrote software that switched its diesels' emissions controls on only when the car detected a laboratory test cycle; on the road, NOx ran up to forty times the US limit, across some eleven million vehicles. The defeat device is Goodhart's Law compiled into an ECU: an algorithm whose entire function is to notice when it is being measured.

How does Goodhart's Law show up in energy KPIs?

The short answer: any energy number attached to a bonus, a compliance filing or a bill starts to drift away from the thing it was supposed to describe. Five scenes, all of them familiar.

The flattering-hour kW/RT. Plant efficiency computed from the 2 p.m. log sheet, when the chillers sit near their sweet spot, and never from the 8 a.m. pulldown or the Saturday part-load slog. Report a monthly average of hand-picked readings and 0.62 kW/RT is easy. The plant did not get more efficient; the sampling did.

The invisible cost of an "efficient" setpoint. Raise chilled-water supply temperature and the kW/RT figure improves immediately. The complaints arrive by WhatsApp and email, which never appear in the energy report. Comfort is the off-balance-sheet liability of an efficiency KPI.

OEE inflated by scheduling. In the factory, the equivalent trick is planning conservatively: schedule less, achieve more of it, and availability looks superb while the plant quietly underproduces.

The benchmark with a trimmed scope. BEI and EUI figures improve wonderfully when the tenant server room, the process line or the basement car park is reclassified out of the denominator's building.

The friendly baseline. M&V has formal wiggle room: ASHRAE Guideline 14 accepts a calibrated baseline with monthly NMBE within ±5% and CV(RMSE) within 15%, meaning two fully compliant baselines can report visibly different savings from the same retrofit. IPMVP's non-routine adjustments are bespoke engineering calculations, each one a lever that can quietly flatter the result when made without disclosure. And ESG figures, assembled downstream from all of the above, are simply the newest bounty on rat tails.

In Malaysia the stakes just went up. The Energy Efficiency and Conservation Act 2024 (Act 859), in force since 1 January 2025, obliges any consumer using 21,600 GJ or more over twelve months to appoint an energy manager, run audits and submit efficiency reports to the Energy Commission. Energy reporting is now a legal target. And TNB's RP4 tariff bills capacity and network charges on the single worst 30-minute demand window of the month (RM89.27/kW for the capacity component under E2, with off-peak maximum demand uncharged for medium-voltage customers), which turns MD into a monthly cash target. Every new target is a new bounty.

The fix is not fewer numbers

Goodhart never said stop measuring. He described what happens when a measure travels alone. The cure is company: instrumentation arranged so that gaming, deliberate or accidental, becomes visible.

Pair every target with a counter-metric that fails when the target is gamed. kW/RT beside space temperatures and the complaint log. OEE beside schedule density. Reported savings beside the weather-normalized raw consumption trend. Keep the interval data flowing next to the reported figure, so the reported figure can always be audited against it. And let anomaly detection be the observer that never sleeps and holds no bonus tied to the KPI: an algorithm flagging that the chiller log and the efficiency report disagree is the modern descendant of the heat-balance test.

Regulators already think this way. Singapore's BCA requires chilled-water plants to carry permanent instrumentation that computes kW/RT within ±5% of the true value, verified by a week-long heat-balance test in which more than 80% of computed balances must land within 5%, with water temperature sensors accurate to ±0.05 °C. The premise is explicit: the measurement layer itself must be independently verifiable, not taken on trust from the party being measured.

This is, frankly, the design brief behind CobiNeural. Continuous sub-metered interval trends sit beside the KPIs they generate (Max Demand, EUI, power factor), so the number can always be interrogated against its own raw data. Condition monitoring on the equipment side can contradict a paper OEE. M&V baselines live in Plan & Verify where everyone can see how they were built, and EECA and ESG reports are generated from the same raw streams. Because it overlays the existing BMS, PLC and SCADA rather than replacing them, the measurement layer stays independent of the systems, and the incentives, it is measuring.

Where the metaphor breaks down

Honesty requires some caveats, because a chiller plant is not a bond market.

Goodhart coined his law for macroeconomic regularities that genuinely collapse when targeted, because markets are made of people who adapt. Physics does not game back. Targeting kW/RT changes nobody's heat-transfer coefficients, and with honest, tamper-resistant metering, a targeted kW/RT remains a perfectly good measure. What degrades in buildings is the reporting layer (which hours get counted, which loads get excluded, which baseline gets picked), not the underlying regularity. The TNB bill remains un-fakeable ground truth at the boundary of the site.

Intent differs too. Volkswagen and Wells Fargo were deliberate fraud. Most plant-room gaming is unconscious selective framing by decent, overloaded engineers who reach for the log sheet that makes the month look survivable. This essay is not accusing its readers of dieselgate; it is pointing out how easily good faith and a lonely KPI produce the same shape of error.

The fix has its own regress. A counter-metric can itself become a target. Anomaly detection is un-gameable only in the sense that it is independent and continuous, not in the sense of being infallible; a determined actor can starve it of sensors or normalize a deviant baseline slowly enough that nothing trips. And some responses to targets are exactly what the target's designer wanted: shifting load off-peak to cut the RP4 capacity charge is the tariff working, not the tariff being gamed. Responding to an incentive is not the same as corrupting a measurement.

Which brings us back to the plant room. Same plant, same 0.62 kW/RT — but now the interval trend, the space temperatures and the complaint log sit on the same dashboard as the headline number. Either the KPI survives the company it keeps, or it doesn't. That is all Goodhart, properly read, ever asked of a measure.

If you want to see what your own KPIs look like sitting next to the raw data that should back them up, request a demo and we'll put both on one screen.